Using deep learning to evaluate peaks in chromatographic data.

Area of Science:

• Analytical Chemistry
• Chemometrics
• Machine Learning

Background:

• Untargeted gas-chromatographic data analysis is time-intensive.
• Previous methods like PARAFAC2-based PARADISe improved efficiency but retained manual steps.
• Identifying suitable peaks for integration from PARAFAC2 components remains a challenge due to peak variability.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for automated peak classification in GC-MS data.
• To address the limitations of linear classifiers in handling variable peak shapes and elution times.
• To further automate the analysis of untargeted GC-MS data, reducing reliance on manual expertise.

Main Methods:

• A CNN classifier was developed to analyze PARAFAC2-resolved components from GC-MS data.
• The CNN's performance was compared against Partial Least Squares regression for Discriminant Analysis (PLS-DA), a shallow artificial neural network, and locally weighted regression.
• Models were trained on over 70,000 elution profile samples from eight GC-MS runs and validated on an independent dataset.

Main Results:

• The CNN classifier achieved an Area Under the Curve (AUC) of 0.95 for peak classification.
• Receiver Operating Characteristic (ROC) curves and misclassification analysis demonstrated superior performance of the CNN over competing models.
• The CNN effectively handled variations in peak shape and elution time, outperforming linear and shallow network approaches.

Conclusions:

• Convolutional neural networks offer a robust and automated solution for peak classification in GC-MS data analysis.
• This AI-driven approach significantly reduces the manual workload and enhances the reliability of untargeted GC-MS data interpretation.
• The developed CNN method provides a convincing means to automatically assess chromatographic data, removing a laborious manual step.